Review agent-written code with impact, provenance, verification evidence, and review state that survives revisions.
~30% lower cost · ~25% faster · without sacrificing task quality
LemonCrow is a local review workspace backed by code intelligence, exact-range reads, bounded tool output, memory, routing, and verification across supported coding-agent hosts.
Real Review Reader · revision 2 · reviewed work preserved while changed and new work returns to the queue.
Install · Review · Results · How it works · Other capabilities · Privacy · Reference
Hosted install (thin client + host integrations, no local LemonCrow server):
curl -fsSL https://github.com/lemoncrow-lab/lemoncrow/releases/latest/download/hosted.sh \
| LEMONCROW_HOSTED_URL=https://your-lemoncrow-server bashFor a local-server production install from this repository, use make prod. It
builds the loopback server into the bundle, starts it on 127.0.0.1:7420 with a
machine credential, and points the same thin client at that local endpoint.
make hosted HOSTED_URL=... exercises the hosted production path from a
checkout. Hosted authentication is owned by Authward; lc auth login performs
the device flow directly against the Authward issuer advertised by the server.
The loopback server is a device-local surface, not a remote deployment target.
Do not place a reverse proxy, Cloudflare Tunnel, SSH port-forward, ngrok, or a
similar tunnel in front of port 7420 to expose Review remotely. Detectable
proxy/forwarding context is refused for local browser trust, but a raw TCP
forward can be indistinguishable from genuine loopback traffic. For remote
Review, use the hosted server path and Authward authentication.
Install once, then open your coding agent in a repository. The LemonCrow SessionStart hook opens the workspace view and syncs only the content the server needs; there is no per-repository local index bootstrap.
Anonymous aggregate telemetry is enabled by default and can be disabled; see Privacy.
Coding agents can produce changes faster than a human can confidently review them. LemonCrow treats human review state as the durable object instead of asking a model to write another review summary.
Start with the local review reader:
lc review --open # local browser review reader
lc review # terminal review summary
lc review --staged # exactly what you are about to commit
Real lc review --open demo · diff → rendered UI preview → compare → review context.
The review workflow is designed to answer:
- what changed and where should I start?
- which changed definitions affect code outside the diff?
- which session or authoring evidence is associated with the change?
- which tests or checks actually ran?
- what have I already reviewed?
- after the author changes the code, which of my prior judgments are still valid?
- which comments still need action?
LemonCrow does not turn those signals into AI APPROVED or SAFE TO MERGE.
Human judgment remains final.
Review marks are attached to content rather than line numbers. If an agent adds lines above something you already reviewed, that work stays reviewed. If the reviewed content changes, LemonCrow puts it back in front of you.
lc review --mark src/auth.py
lc review --comment "ttl is hardcoded" --on src/auth.py:L81
# author or agent revises the change
lc review --since-my-review
lc review --feedbackComments are re-anchored only when the new location is unambiguous. Otherwise they are reported as orphaned instead of being silently moved.
lc review also exposes impact outside the patch, review ordering, provenance,
and recorded verification evidence. Missing evidence is reported as unknown;
LemonCrow does not synthesize a passing test status.
See the 10-minute review walkthrough or the CLI reference.
The review workflow runs on the same code-intelligence runtime used by coding agents. The numbers below measure the runtime on complete tasks. They are fixed results from pinned benchmark runs, not live counters. Full methodology and raw-run references are in BENCHMARKS.md.
~30% lower cost · ~25% faster · without sacrificing task quality
SWE-bench Verified, same model and harness: 29.5% lower cost, 23.7% less wall-clock time, and 92.8% resolved vs 80.8% baseline.
| Benchmark | Baseline correct | LemonCrow correct | Baseline cost | LemonCrow cost | Cost delta | SWE-bench Verified, same model and harness: 29.5% lower cost, 23.7% less wall-clock time, and 92.8% resolved vs 80.8% baseline.
| Benchmark | Baseline correct | LemonCrow correct | Baseline cost | LemonCrow cost | Cost delta |
|---|---|---|---|---|---|
| SWE-bench Verified, 50 tasks × 5 reps | 80.8% | 92.8% | $234.84 | $165.45 | 29.5% cheaper |
| SWE-bench Lite, 10 tasks × 5 reps | 98.0% | 96.0% | $19.83 | $17.51 | 11.7% cheaper |
| SWE-bench Pro, 10 tasks × 5 reps | 88.0% | 90.0% | $39.01 | $30.61 | 21.5% cheaper |
| Terminal-Bench 2.1, 89 tasks × 5 reps, Opus 4.8 | 78.9% | 78.9% | $73.75 | $61.98 | 16.0% cheaper |
The table includes the SWE-bench Lite regression: LemonCrow was cheaper on that run but scored 2 percentage points lower. The benchmark set is not filtered to keep only favorable results.
On SWE-bench Verified, the same pinned run used 37.7% fewer turns, 23.7% less wall-clock time, 37.8% fewer tool calls, and 27.9% fewer output tokens.
For retrieval and scale, the benchmark suite also includes ~7,200 query/answer pairs across 14 repositories. LemonCrow measured 0.727 semantic retrieval MRR versus 0.376 for ripgrep, with a 390 ms semantic-search p95. A cold lexical index of Linux kernel core (1.24M symbols, 4.5M lines) measured 179 seconds. See indexing and retrieval results for the full comparison and caveats.
LemonCrow keeps your existing coding agent and changes the working set around it. The runtime is local and operates across four stages:
| Stage | What LemonCrow does |
|---|---|
| Find | Rank symbols, definitions, callers, callees, usages, and exact source ranges before broad file exploration. |
| Read | Return an outline or requested ranges; cap noisy command and web output with recoverable spill files. |
| Carry | Preserve useful task state through memory, deduplication, compaction manifests, and handover packets. |
| Verify | Record verification evidence and notice code changes that have no corresponding checks. |
LemonGraph — the local code graph used for search, impact, and review context.
On Claude Code, lc init can replace equivalent built-in exploration tools with
LemonCrow's grounded tools. Other hosts use the strongest controls they expose;
capability depth varies by host.
| LemonCrow tool | Purpose |
|---|---|
code_search |
Ranked code search with definitions, callers/callees, usages, and source ranges. |
read |
Outlines or exact ranges instead of full-file reads by default. |
edit |
Verified multi-file edits. |
bash |
Bounded structured command output with recoverable spill files. |
web_fetch |
Clean page extraction instead of raw HTML. |
The point is not to optimize one isolated retrieval call. A coding task is a loop: find → read → act → carry → verify. LemonCrow measures the completed task because saving tokens in one step is not useful if the agent buys them back in extra turns later.
Architecture details: docs/reference/architecture.md.
Supported integrations include Claude Code, Codex CLI, Cursor, opencode, LemonCode, Pi, Copilot, Copilot CLI, Hermes Agent, and Antigravity. They do not all expose the same hooks or enforcement controls.
See all host integrations and the host capability matrix before depending on a host-specific lifecycle or verification feature.
These are useful, but they are not the core review workflow:
- Remote MCP:
lc mcp serveexposes the current workspace through an OAuth-protected remote MCP endpoint for compatible clients. - Session replay:
lc session statsandlc session replayinspect recorded agent sessions locally without rerunning the model. - Usage accounting:
lc usagereports recorded usage by host, model, project, and day without treating unknown pricing as zero. - Bring your own model:
lc modelregisters and probes OpenAI-compatible endpoints such as vLLM, Ollama, LM Studio, or an internal gateway. - Agents and skills: packaged agent modes and optional skills live under
integrations/.
Full commands and flags live in the CLI reference.
LemonCrow's core runtime is local. Indexing, search, edits, memory, review state, and reports stay on your machine. No LemonCrow account is required.
Anonymous aggregate telemetry is on by default. It contains aggregate counts, bucketed durations, dollar estimates, hashed install/session identifiers, version, host source, retrieval domain, and a timestamp. It does not contain source code, prompts, repository paths, file paths, or symbol names.
Turn remote telemetry off with:
lc telemetry remote offor set DO_NOT_TRACK=1 / LEMONCROW_TELEMETRY=off.
Commands you explicitly run may still use the network: for example lc update
checks GitHub Releases, configured model providers receive model requests, and
optional dependency bootstrap may fetch upstream artifacts. See the complete
privacy and network behavior document.
- Benchmark results are measured on the published task sets; they are not a guarantee for your repository or model.
- Host capabilities differ. Features that depend on lifecycle hooks or live session capture may degrade to imported or manual evidence on some hosts.
- Linux and macOS are the primary supported operating systems. Windows support is partial and not currently validated.
- LemonCrow does not run or pay for your model provider; you configure the provider or local endpoint you want to use.
- This repository is the local open-source runtime. Hosted/enterprise deployments use a remote LemonCrow server plus Authward and commercial entitlement; none of that is required for local use.
I built LemonCrow after using coding agents heavily and seeing two bottlenecks move into focus. First, agents repeatedly spent context and tool calls rediscovering the same code. Once that became cheaper, the harder bottleneck was human attention: the agent could change code faster than I could confidently review it.
My background at Google included performance optimization and cost work, so the runtime is measured at the whole-task level rather than by claiming a win on one isolated hop. But the product direction is review-first: reduce how much a human has to reread without outsourcing the engineering judgment.
- Installation
- Review walkthrough
- CLI reference
- Architecture
- Benchmarks
- Full benchmark results
- Host integrations
- Host capability matrix
- Privacy & network behavior
- Troubleshooting
Uninstall LemonCrow and its host integrations while preserving local data by default:
bash scripts/uninstall.shRemove LemonCrow-managed local state too:
bash scripts/uninstall.sh --purgePreview with --dry-run.
git clone https://github.com/lemoncrow-lab/lemoncrow
cd lemoncrow
bash scripts/local.shSee Installation for development setup details.
Apache-2.0. The complete local runtime is open source.
Hosted LemonCrow is a separate enterprise product under development. It is not part of this repository and is not required for local use. If you want to discuss an enterprise pilot or design-partner deployment, contact [email protected].